OpenAI

Rapidly scaling online storage to serve over 1 billion ChatGPT users (Part II)

OpenAI

(LLM summary failed, please supplement manually later)

Securing our frontier model training environment

OpenAI

(LLM summary failed, please supplement manually later)

Anthropic News

Announcements

Anthropic

This is a summary of announcements from AI company Anthropic, a R&D institution focused on the AI safety field whose core R&D goal is to build highly reliable, interpretable, controllable AI systems. Its R&D path prioritizes safety as a core leading principle, aiming to avoid safety risks such as uncontrollability and black box issues in AI development from the underlying technology layer, laying a solid safety foundation for the implementation of artificial general intelligence.

Investing in the public benefit

Anthropic

AI company Anthropic has launched a public benefit-linked equity arrangement: part of the company’s shares will be allocated to the newly established “Anthropic Public Benefit Fund”, which adopts a special structure where the share proportion it holds will increase synchronously as the company scales and its valuation grows. This mechanism directly links corporate growth dividends with public benefits, and is an innovative practice for technology companies to explore feeding development gains back to public welfare.

Google DeepMind

Beyond tree cover: mapping the world’s forest types

Google DeepMind

Beyond tree cover: mapping the world’s forest types breaks through the traditional approach of only counting forest coverage, and for the first time achieves fine-grained classified mapping of forest types for each terrestrial plot unit at global scale with 10-meter resolution. The results can support accurate accounting of global forest carbon sinks, biodiversity conservation, dynamic management of forest resources and climate change response research, filling the gap in high-resolution mapping in this field.

Accelerating the search for antimicrobial peptides with Google Antigravity

Google DeepMind

Bacterial antibiotic resistance is currently a severe issue, and the development of new antimicrobial peptides is an important response path, but traditional candidate molecule screening efficiency is low. This research relies on Google’s “Antigravity Science” tool to optimize the antimicrobial peptide screening process, which can greatly improve the efficiency of discovering new antimicrobial peptides, not only reducing costs and increasing efficiency for pre-research of new antimicrobial peptide drugs, but also providing a new technical idea for high-throughput screening of other functional peptides.

Hugging Face Blog

LitPivot: Developing research ideas through dynamic contextualization within the literature landscape

Hugging Face

This research proposes LitPivot, an auxiliary tool for scientific research topic selection, whose core method is to realize dynamic contextual association in the global literature graph, breaking through the limitations of traditional keyword retrieval, intelligently identifying cross-domain research gaps, and outputting adapted innovative topic ideas combined with the researcher’s own foundation. Actual testing shows that the research ideas it assists in producing are significantly better than traditional literature tools in terms of innovation and implementation feasibility.

Does Your Model Run on AWS Trainium? Look It Up

Hugging Face

This article addresses the pain point of high cost for checking model migration compatibility for AWS’s self-developed Trainium training chip, and launches an open-source automatic query tool that covers mainstream tasks such as LLM, CV, multimodal, and mainstream deep learning frameworks such as PyTorch. It can automatically detect operator support and provide adaptation solutions, with a compatibility judgment accuracy rate of over 95%, and troubleshooting efficiency 80% higher than manual work, greatly lowering the migration threshold for users.

Lil’Log

Harness Engineering for Self-Improvement

Lilian Weng

The concept of Recursive Self-Improvement (RSI) was first proposed by I.J. Good in 1965, referring to the operating logic of superintelligence that can surpass all human intelligent activities and iteratively design better systems; in 2008, Eliezer Yudkowsky clarified that its core is the feedback loop where AI relies on existing intelligence to optimize its own cognitive architecture. RSI in the current AI context includes both the path where the model directly rewrites its own weights, and can also broadly refer to the behavior of the model optimizing the training pipeline.

量子位 (QbitAI)

iFlytek launches the Star TaaS platform, making Token a real productivity tool for thousands of industries

QbitAI

At the iFlytek 1024 Developer Festival in October 2026, multiple AI achievements were released: the end-cloud integrated Spark X2.6 large model has been expanded to a 1 million token context window, end-cloud routing accuracy has been increased to 96.1%, and a 4B end-side model has been open-sourced simultaneously; the SparkAudio 1.0 speech foundation model trained entirely on domestic computing power was launched for the first time, with greatly optimized comprehensive performance; tools such as the Star TaaS platform and Agent OS were also launched, the platform currently has more than 11.87 million developers, with the goal of promoting Token to become productivity for thousands of industries.

Enterprise AI needs two world models

QbitAI

Currently, enterprise AI is generally stuck in an implementation dilemma: 95% of generative AI pilots have no clear profit and loss returns, and applications mostly stay at the individual employee level, without being converted into actual productivity. Fan Ling, founder of Tezign, proposed a solution at the WTCC forum: enterprises do not need to blindly upgrade general large models, but should build two sets of world models: a subjective model relying on external user data to understand customer preferences, and an enterprise model relying on internal data to adapt to internal process rules.

Run a 34B real-time world model on a single 5090, domestic Chinese team wins NVIDIA award!

QbitAI

Chongqing Dawar Technology’s real-time world model Scenes AI won the Excellence Award at the mid-term showcase of the 2026 NVIDIA Startup Exhibition, where more than 580 enterprises participated. This model does not need to predict the next frame, but maintains independently readable and writable world state variables and only updates the changed parts. The input is real footage from ordinary cameras, and a single RTX 5090 can run the 34B parameter model, which can achieve effects such as real-time style conversion while retaining the real spatial structure.